Zig 2026上半年：构建系统分离、@bitCast 新语义、SPIR-V 后端复苏、增量编译提速 10 倍

Now that there is a separate process for users’ build.zig scripts and the build system itself, it makes sense for that to be the place that package management logic lives.

I moved these subcommands to the maker process:

zig build

zig fetch

zig init

zig libc

This means that large parts of what used to be included in the compiler executable are now shipped in source form instead, including:

package fetching logic

HTTP client and networking

TLS (Transport Layer Security) and associated crypto

Git protocol

xz, gzip, zstd, flate, zip

parsing, validation, and otherwise dealing with build.zig.zon files

Consequently, this functionality can now be patched without rebuilding the compiler, making it easier for users and contributors to tinker.

Furthermore, it means that package management in zig now has safety checks enabled when doing networking, since the maker executable is compiled in ReleaseSafe mode. Plus, all the crypto used for networking and file hashing can now take advantage of special CPU instructions available on the host, even the ones that are too rare to normally depend on when distributing software. We can have AOT cake and eat JIT, too!

My original motivation for doing this was in relation to exposing a build server protocol in order to unblock ZLS after maker/configurer process separation made breaking changes to the --build-runner override flag.

Originally, the process tree looked like this:

zig build (the zig compiler + package manager) └─ builder (the user's build.zig logic + build system implementation)

The process separation changeset made it look like this instead:

zig build (the zig compiler + package manager) ├─ configurer (the user's build.zig logic) └─ maker (build system)

At this point, consider a long-running zig build --watch process, watching files and rebuilding on source code changes. If any changes to build.zig are detected, or any files observed during execution of that logic, it means configurer needs to be rerun, meaning that maker process must exit to give zig build a chance to repeat the package management logic.

Now, after the changes described in this devlog entry, it looks like this:

zig build (the zig compiler) └─ maker (build system + package manager) └─ configurer (the user's build.zig logic)

Thus, when configuration needs to be rerun, maker process can continue to live because it is the parent process rather than a sibling. In terms of the upcoming build server, it means avoiding an awkward situation where the server has to exit and the client has to reconnect, rather than simply informing the client of a configuration change.

This is almost entirely a non-breaking change, but there are some observable differences:

Zig executable binary size: shrinks 4% from 14.1 to 13.5 MiB (no LLVM, ReleaseSmall)

--maker-opt flag is replaced by ZIG_DEBUG_MAKER environment variable

--zig-lib-dir flag is replaced by ZIG_LIB_DIR environment variable

The follow-up issues to this changeset are the main blockers until we tag Zig 0.17.0:

build server protocol MVP (needed to unblock ZLS)

introduce the concept of adding path dependencies of the build script itself

make zig build --watch detect modifications to the build script and rerun itself

different cwd causes build script cache miss

I have two conferences coming up in July and I need to work on my talks, so being realistic, I don’t think I will have time to wrap these up until early August. Contributions welcome, of course.

Big thanks to Techatrix from the ZLS team for reaching out and working with me on the build server protocol! They are seeking sponsorship, by the way.

There’s quite a bit to cover. The SPIR-V backend had bitrotted in a number of places after the recent compiler changes, so I spent the past several weeks dragging it into a better state.

@SpirvType

SPIR-V has a handful of types that couldn’t be expressed in Zig’s type system. The new @SpirvType builtin has been introduced to address the longest-standing blocker for writing shaders. See #20550, #23326 and #35461 to trace the background.

const Sampler = @SpirvType(.sampler); const Image = @SpirvType(.{ .image = .{ .usage = .{ .sampled = u32 }, .format = .unknown, .dim = .@"2d", .depth = .unknown, .arrayed = false, .multisampled = false, .access = .unknown, } }); const SampledImage = @SpirvType(.{ .sampled_image = Image }); const RuntimeArray = @SpirvType(.{ .runtime_array = u32 }); const sampled_image = @extern(*addrspace(.constant) const SampledImage, .{ .name = "sampled_image", .decoration = .{ .descriptor = .{ .set = 0, .binding = 1 } }, });

const Sampler = @SpirvType(.sampler);
const Image = @SpirvType(.{ .image = .{
    .usage = .{ .sampled = u32 },
    .format = .unknown,
    .dim = .@"2d",
    .depth = .unknown,
    .arrayed = false,
    .multisampled = false,
    .access = .unknown,
} });
const SampledImage = @SpirvType(.{ .sampled_image = Image });
const RuntimeArray = @SpirvType(.{ .runtime_array = u32 });
const sampled_image = @extern(*addrspace(.constant) const SampledImage, .{
    .name = "sampled_image",
    .decoration = .{ .descriptor = .{ .set = 0, .binding = 1 } },
});

Execution Mode on the Calling Convention

Execution mode info (workgroup size, fragment origin, etc.) is now carried by the calling convention instead of being emitted via inline assembly OpExecutionMode. The old std.gpu.executionMode() helper is gone, and the SPIR-V assembler now rejects manual OpExecutionMode instructions. Two new calling conventions, spirv_task and spirv_mesh, were also added for mesh shading pipelines.

export fn vert() callconv(.spirv_vertex) void {} export fn frag() callconv(.{ .spirv_fragment = .{ .depth_assumption = .greater } }) void {} export fn comp() callconv(.{ .spirv_kernel = .{ .x = 8, .y = 8, .z = 1 } }) void {} export fn task() callconv(.{ .spirv_task = .{ .x = 1, .y = 1, .z = 1 } }) void {} export fn mesh() callconv(.{ .spirv_mesh = .{ .stage_output = .output_lines, .max_primitives = 1, .max_vertices = 2 } }) void {}

Capabilities and Extensions from CPU Features

Capabilities and extensions used to be emitted ad hoc by codegen or via inline assembly. They’re now driven entirely by the CPU feature set like other targets, with dependency chains extracted from SPIRV-Headers (excluding external vendors for now), and the assembler now rejects any attempt to emit OpCapability or OpExtension directly.

Multi-Threaded Codegen

From day one, the SPIR-V backend ran codegen single-threaded inside the linker thread. Each codegen job now produces an Mir value just like every other self-hosted backend, and gets scheduled on the compiler’s thread pool.

The same change brought back two ISel passes that had been removed during earlier refactors: dedup_types (which merges equivalent type instructions) and prune_unused (which strips dead code from the final module). These had originally been deleted back when codegen was single-threaded.

Object File Linking

.spv files are now recognised as object files. You can compile multiple .zig files (or external .spv objects) and have the SPIR-V linker stitch them into a single module.

Tens of bugs have also been fixed along the way with a nearly 10% increase in total passing behavior tests (49% now) on the spirv64-vulkan target, std.gpu was renamed to std.spirv and the SPIR-V backend is meaningfully more useful than it was a month ago, but there’s still a long way to go. Plenty of behavior tests remain skipped on SPIR-V. That said, if you’ve been on the fence about trying Zig for shaders or compute kernels, this is a good time to give it a shot. Bug reports are very welcome on Codeberg. Happy hacking!

export fn vert() callconv(.spirv_vertex) void {}
export fn frag() callconv(.{ .spirv_fragment = .{ .depth_assumption = .greater } }) void {}
export fn comp() callconv(.{ .spirv_kernel = .{ .x = 8, .y = 8, .z = 1 } }) void {}
export fn task() callconv(.{ .spirv_task = .{ .x = 1, .y = 1, .z = 1 } }) void {}
export fn mesh() callconv(.{ .spirv_mesh = .{ .stage_output = .output_lines, .max_primitives = 1, .max_vertices = 2 } }) void {}

A few weeks ago, I began working on a branch implementing an improvement to the LLVM backend which had been planned for a long time. This ended up snowballing into a bigger change which implemented a few language proposals you might be interested to hear about.

LLVM Backend Integer Lowering

Zig has always lowered arbitrary bit-width integer types (e.g. u4, i13, u40) directly to LLVM IR’s bit-int types (i4, i13, i40). However, we’ve known for a long time that this lowering is not optimal, because LLVM’s documented semantics for representing these types in memory are unnecessarily restrictive to the optimizer. Perhaps more importantly, because Clang never emits LLVM IR like this, these code paths in LLVM have never been properly tested, and so are poorly supported in practice—over the past few years, we have observed many instances of trivial optimizations being missed and even straight-up miscompilations.

So, the original goal of the PR was to only use these bit-int types when manipulating values in SSA form, and to zero- or sign-extend them to ABI-sized types (i8, i16, i32, etc) when storing them in memory. This should be well-supported, not least because it matches how Clang lowers C’s _BitInt(N)!

That change was actually fairly straightforward, but I hit one issue which led me down a bit of a rabbit-hole.

The Problem with @bitCast

@bitCast is an interesting builtin. In the past, it was defined as being equivalent to the following sequence of operations:

Take a pointer to the operand value

Cast it to a pointer to the destination type

Load from that pointer

In other words, it was essentially syntax sugar for reinterpreting bytes of memory. However, over time, we diverged from this definition—for instance, it became allowed to use @bitCast to reinterpret a [3]u8 as a u24, even though on most targets @sizeOf(u24) is greater than @sizeOf([3]u8) so the above definition would invoke Illegal Behavior.

Up to now, the LLVM backend had implemented these underspecified semantics for the @bitCast builtin. However, because that definition involved reinterpreting memory, changing how we store integer types in memory ended up impacting the implementation of @bitCast, and introducing Illegal Behavior which led to crashes in the compiler test suite.

The easiest solution to this would probably have been to implement logic in the LLVM backend to approximately match the old behavior. I instead opted for a better solution—implement a new definition of @bitCast.

Redefining @bitCast

In 2024, Jacob Young wrote up language proposal #19755 which aimed to solve the problems with @bitCast by precisely specifying a new set of semantics for it. This proposal was accepted shortly after it was submitted, and in fact, the semantics it details are already implemented by the self-hosted x86_64 backend! So to solve the LLVM backend’s problems, I didn’t necessarily need to match the old @bitCast semantics—instead, this seemed like a good time to finally get the new semantics implemented everywhere.

As an aside, another advantage to doing this is that we could take advantage of the compiler’s Legalize pass, which takes difficult-to-lower operations and rewrites them in terms of simpler operations, so that compiler backends only need to support those simple operations. Legalize already had functionality, used by the self-hosted x86_64 backend, which converted complex @bitCast operations into simpler ones, and it could be easily adapted to aid the other compiler backends too (mainly the LLVM and C backends)—but only if they implemented the new semantics.

Regardless, the point is, I set out on a side quest (which ended up being harder than the original quest) to implement these new semantics throughout the compiler. This includes not only the LLVM and C backends, but also comptime execution—after all, Zig allows you to do almost any operation at comptime, @bitCast included! Because the new semantics are meaningfully different from the old (more on this later), I also had to audit a lot of uses of @bitCast across the standard library, compiler, and supporting libraries (e.g. compiler_rt). But after a few mostly-painless fixes for CI failures, I was able to finally get my PR green, and landed it in master yesterday (closing a good few issues in the process!).

The New @bitCast Semantics

Now that we’ve gotten through all of the background, it’s finally time for me to actually explain new @bitCast behavior. Instead of being based on reinterpreting bytes in memory like before, the builtin is now defined in terms of the bits which logically represent a type.

Every type which supports @bitCast has a “logical bit layout”—a representation of that type as an ordered sequence of bits. For instance, u5 is composed of 5 logical bits, which we order from least-significant to most-significant. [2]u5 is composed of 10 logical bits—the 5 from the first element, followed by the 5 from the second element. The new definition of @bitCast is that it reinterprets the logical bits of one type as the logical bits of a different type.

The simplest example is to take an unsigned integer, say a u8, and convert it to a signed integer of the same size, in this case i8. This operation does exactly what you’d expect—the bits are unchanged, and we just reinterpret the most-significant bit as a sign bit. Also unchanged are the semantics of @bitCast between an integer type and a packed struct/packed union type.

The place where the new semantics differ from the old is when you get aggregate types (arrays and vectors) involved.

Consider, for instance, bitcasting a [2]u8 to a u16. Under the old semantics, the result of this operation depends on the target endian: on big-endian targets, the first array element became the 8 most significant bits, whereas on little-endian targets, the first array element became the 8 least significant bits. Under the new semantics, because we only care about logical bit representation (which is endian-agnostic), the operation behaves identically on every target: the first array element becomes the 8 least significant bits. As a general rule, the new semantics tend to match the behavior of the old semantics on little-endian targets.

This definition also allows for some weirder operations, such as converting [2]u3 to @Vector(3, u2):

test "bitcast [2]u3 to @Vector(3, u2)" { const arr: [2]u3 = .{ 0b001, 0b011 }; const vec: @Vector(3, u2) = @bitCast(arr);

// Concatenate all bits of arr starting with the least-significant bit of arr[0] to find the // logical bit sequence, then read off 2-bit chunks from it to get the elements of the resulting // vector value vec. // // arr[0] arr[1] // 0b001 0b011 // ------------- ------------- // 1 0 0 1 1 0 // -------- -------- -------- // 0b01 0b10 0b01 // vec[0] vec[1] vec[2]

try expect(vec[0] == 0b01); try expect(vec[1] == 0b10); try expect(vec[2] == 0b01); } const expect = @import("std").testing.expect;

This kind of operation isn’t very useful most of the time, but it’s there if you need it! For instance, perhaps you want to deconstruct an integer into a vector of individual bits to operate on—that can now be done by a @bitCast to @Vector(n, u1).

While doing all of this stuff, I also implemented a couple of smaller accepted proposals—I won’t detail them here, but you can take a look at the issues if you’re interested:

Disallow @bitCast to/from vectors of pointers (#18936)

Allow @bitCast on enums (part of #35602)

Of course, all of these changed semantics will be explained in the 0.17.0 release notes (hopefully a bit more concisely than what I managed here!), and suggested migration steps outlined.

LLVM Backend Performance

On a final note, I just wanted to mention that the original motivation for this branch—changing how the LLVM backend lowers non-ABI integer types—was demonstrably successful at restoring missed optimizations. In fact, the Zig compiler itself—despite not making heavy use of arbitrary bit width integers internally!—saw around 5% performance improvements from the better optimization. This means you might have some minor runtime performance gains to look forward to in 0.17.0!

Thanks for reading, I hope this was interesting to some of you. Happy hacking!

test "bitcast [2]u3 to @Vector(3, u2)" {
    const arr: [2]u3 = .{ 0b001, 0b011 };
    const vec: @Vector(3, u2) = @bitCast(arr);

    // 将 `arr` 的所有位从 `arr[0]` 的最低有效位开始拼接，得到逻辑位序列，
    // 然后从中读取 2 位的数据块，得到结果向量值 `vec` 的元素。
    //
    // arr[0]   arr[1]
    // 0b001    0b011
    // ------------- -------------
    // 1 0 0     1 1 0
    // -------- -------- --------
    // 0b01     0b10     0b01
    // vec[0]   vec[1]   vec[2]

    try expect(vec[0] == 0b01);
    try expect(vec[1] == 0b10);
    try expect(vec[2] == 0b01);
}
const expect = @import("std").testing.expect;

I’ve spent the past few weeks working on our new ELF linker which debuted in Zig 0.16.0. At the time of the 0.16.0 release, this linker implementation was in its fairly early stages, and only really supported linking Zig-only code without any external libraries (even libc)—hence why it was (and still is) disabled by default (it can be enabled with -fnew-linker). However, quite a lot of progress has been made since that initial release!

Here’s a nice milestone—as of my latest PR, the new ELF linker is capable of building the self-hosted Zig compiler with LLVM and LLD libraries enabled, a task which requires quite a few features under the hood.

[mlugg@nebula master]$ # Build the Zig compiler using the new linker: [mlugg@nebula master]$ zig build -Dno-lib -Dnew-linker -Denable-llvm [mlugg@nebula master]$ # Use that compiler to build something with LLVM and LLD: [mlugg@nebula master]$ ./zig-out/bin/zig build-exe ~/hello.zig -fllvm -flld [mlugg@nebula master]$ ./hello Hello, World! [mlugg@nebula master]$

Of course, an ELF linker isn’t necessarily the most exciting thing in the world, which is why the headline feature of this new linker is its support for fast incremental compilation. After the recent enhancements, it is now possible (on x86_64 Linux) to perform incremental rebuilds while linking external libraries, C sources, etc—without any additional performance overhead! Here’s a clip of me trying it out on Andrew’s Tetris clone:

A few silly changes to Andrew’s Tetris clone being built in around 30ms each.

[mlugg@nebula master]$ # 使用新链接器构建 Zig 编译器：
[mlugg@nebula master]$ zig build -Dno-lib -Dnew-linker -Denable-llvm
[mlugg@nebula master]$ # 使用该编译器构建带有 LLVM 和 LLD 的程序：
[mlugg@nebula master]$ ./zig-out/bin/zig build-exe ~/hello.zig -fllvm -flld
[mlugg@nebula master]$ ./hello
Hello, World!
[mlugg@nebula master]$

Oh, and fast incremental rebuilds also work nicely on the Zig compiler itself:

[mlugg@nebula master]$ zig build -Dno-lib -Denable-llvm -fincremental --watch Build Summary: 4/4 steps succeeded install success └─ compile exe zig Debug native success 36s

Build Summary: 4/4 steps succeeded install success └─ compile exe zig Debug native success 244ms

Build Summary: 4/4 steps succeeded install success └─ compile exe zig Debug native success 228ms

Build Summary: 4/4 steps succeeded install success └─ compile exe zig Debug native success 288ms

Build Summary: 4/4 steps succeeded install success └─ compile exe zig Debug native success 283ms

The biggest missing feature of this linker implementation right now is that it still does not yet support generating DWARF debug information for Zig code—that’s definitely my next priority. But even without that support, it’s amazing just how useful instant rebuilds can be, for example in any situation where you’re doing a lot of print debugging.

If you’re using the master branch of Zig and you’re on x86_64 Linux, consider trying out incremental compilation with the new ELF linker if it previously wasn’t working with your project! I expect many codebases to already work great with it, unlocking the ability to rebuild your project in milliseconds. Of course, if you come across any bugs, please do open an issue.

And if you’re currently sticking to tagged releases of Zig, don’t worry—as Andrew mentioned in his last devlog, Zig 0.17.0 is just around the corner, so it won’t be long before you can try this too!

[mlugg@nebula master]$ zig build -Dno-lib -Denable-llvm -fincremental --watch
Build Summary: 4/4 steps succeeded
install success
└─ compile exe zig Debug native success 36s

Build Summary: 4/4 steps succeeded
install success
└─ compile exe zig Debug native success 244ms

Build Summary: 4/4 steps succeeded
install success
└─ compile exe zig Debug native success 228ms

Build Summary: 4/4 steps succeeded
install success
└─ compile exe zig Debug native success 288ms

Build Summary: 4/4 steps succeeded
install success
└─ compile exe zig Debug native success 283ms

Big branch just landed: separate the maker process from the configurer process

This devlog entry is essentially a preview of the upcoming release notes, but serves as an advanced notice to those who want to help test out the new features and provide feedback that will guide the Zig project moving forward.

Before, build.zig files plus the build system implementation were all compiled into one bloated process, in Debug mode. After build.zig logic finished constructing a build graph in memory, the “build runner” code executed it.

Now, build.zig files are compiled into a small process (the “configurer”) in debug mode. After this logic finishes constructing a build graph in memory, it is serialized to a binary configuration file. The parent zig build process is aware of this file and caches it for next time. While waiting for all that, it asynchronously compiles the build graph execution process (the “maker”) in release mode. Once the configuration file is available and the maker process is finished compiling, the maker process is executed, passing it the configuration file. The maker process only needs to be compiled once per zig version thanks to the global cache. The maker process then executes the build graph, which is contained within the serialized configuration file.

The primary motivation of this change was to make zig build faster, in three ways:

Only the user’s build.zig logic will be compiled with each change, rather than the entire build system along with it. This is starting to become more valuable now that we have introduced --watch, --fuzz and --webui. The build system can grow more features without making zig build take longer.

Now the build system can skip rerunning the build.zig logic entirely when it knows nothing will change, for example if you add -freference-trace to your zig build command line, it now avoids re-running your build.zig logic redundantly, using the same configuration as last time.

Now the process that actually executes the build graph is compiled with optimizations enabled.

To demonstrate points 2 and 3, here is the difference between running zig build --help before and after:

Benchmark 1 (34 runs): master/zig build -h measurement mean ± σ min … max outliers delta wall_time 150ms ± 5.52ms 145ms … 165ms 4 (12%) 0% peak_rss 84.8MB ± 275KB 84.2MB … 85.1MB 0 ( 0%) 0% cpu_cycles 593M ± 4.01M 588M … 608M 2 ( 6%) 0% instructions 995M ± 52.5K 995M … 995M 0 ( 0%) 0% cache_references 25.8M ± 165K 25.4M … 26.1M 0 ( 0%) 0% cache_misses 651K ± 20.1K 619K … 697K 0 ( 0%) 0% branch_misses 918K ± 7.44K 906K … 935K 0 ( 0%) 0% Benchmark 2 (348 runs): branch/zig build -h measurement mean ± σ min … max outliers delta wall_time 14.3ms ± 744us 13.2ms … 23.3ms 8 ( 2%) ⚡- 90.4% ± 0.4% peak_rss 78.5MB ± 562KB 77.1MB … 81.4MB 7 ( 2%) ⚡- 7.4% ± 0.2% cpu_cycles 24.1M ± 821K 22.8M … 27.1M 3 ( 1%) ⚡- 95.9% ± 0.1% instructions 43.7M ± 23.8K 43.7M … 43.8M 56 (16%) ⚡- 95.6% ± 0.0% cache_references 1.46M ± 14.6K 1.40M … 1.50M 19 ( 5%) ⚡- 94.3% ± 0.1% cache_misses 142K ± 4.87K 127K … 157K 2 ( 1%) ⚡- 78.1% ± 0.4% branch_misses 126K ± 1.37K 120K … 129K 12 ( 3%) ⚡- 86.3% ± 0.1%

It’s dramatic because before, build.zig logic was being executed with each zig build command, but now, the build system uses the cached, serialized configuration instead.

Aside from performance, I expect third-party tooling such as ZLS to benefit from consuming the serialized configuration file rather than maintaining a fork of the build runner.

Benchmark 1 (34 runs): master/zig build -h
  wall_time     150ms ± 5.52ms    145ms … 165ms    4 (12%)     0%
  peak_rss      84.8MB ± 275KB    84.2MB … 85.1MB  0 ( 0%)     0%
  cpu_cycles    593M  ± 4.01M     588M  … 608M     2 ( 6%)     0%
  instructions  995M  ± 52.5K     995M  … 995M     0 ( 0%)     0%
  cache_references 25.8M ± 165K   25.4M … 26.1M    0 ( 0%)     0%
  cache_misses  651K  ± 20.1K     619K  … 697K     0 ( 0%)     0%
  branch_misses 918K  ± 7.44K     906K  … 935K     0 ( 0%)     0%

Benchmark 2 (348 runs): branch/zig build -h
  wall_time     14.3ms ± 744us    13.2ms … 23.3ms  8 ( 2%)   ⚡- 90.4% ± 0.4%
  peak_rss      78.5MB ± 562KB    77.1MB … 81.4MB  7 ( 2%)   ⚡- 7.4% ± 0.2%
  cpu_cycles    24.1M  ± 821K     22.8M  … 27.1M   3 ( 1%)   ⚡- 95.9% ± 0.1%
  instructions  43.7M  ± 23.8K    43.7M  … 43.8M  56 (16%)   ⚡- 95.6% ± 0.0%
  cache_references 1.46M ± 14.6K  1.40M  … 1.50M  19 ( 5%)   ⚡- 94.3% ± 0.1%
  cache_misses  142K  ± 4.87K     127K  … 157K     2 ( 1%)   ⚡- 78.1% ± 0.4%
  branch_misses 126K  ± 1.37K     120K  … 129K    12 ( 3%)   ⚡- 86.3% ± 0.1%

This changeset heavily reworks the internal mechanism of the zig build system, however, it is mostly non-breaking from an API perspective, with the exceptions noted in the PR linked above.

For most people I’m guessing this is the main breaking change they’ll hit:

if (b.args) |args| { run_cmd.addArgs(args); }

⬇️

run_cmd.addPassthruArgs();

This removes a capability from build scripts since they can no longer observe those arguments. In exchange, it means that when changing those arguments, build scripts no longer must be rebuilt from source.

If you’re someone who wants to influence the direction of Zig, this is a good time to upgrade your projects to the development version and try out these changes. We’ll be releasing 0.17.0 within a couple weeks from now. However, if you don’t have time, and you find out that 0.17.0 broke your build, don’t worry, there will be plenty of opportunity to get fixes in for the 0.17.1 tag as well.

if (b.args) |args| {
    run_cmd.addArgs(args);
}

run_cmd.addPassthruArgs();

I’ve been spending a bit of time working on personal projects after merging my type resolution changes last month, but I did find the time recently to make some improvements to the LLVM codegen backend. This involved a few different enhancements with various goals, but one nice user-facing change was that I managed to get incremental compilation working with the LLVM backend.

Sadly this can’t do anything to speed up the dreaded LLVM Emit Object: that time is entirely down to LLVM. However, what incremental compilation does help with is minimizing the time spent in the actual Zig compiler code, which means that if your code has compile errors (so “LLVM Emit Object” will be skipped), you’ll usually get those errors very quickly. (Of course, it does still give you a slight speed-up in successful builds too.)

This support is available in master branch builds right now, and will be in the 0.16.0 release (which we’ll be tagging very soon).

For anyone who still hasn’t tried it, especially if you’re using Zig’s master branch, please do try out incremental compilation by passing -fincremental --watch to zig build! The Zig core team have benefited from incremental compilation in our workflows for a good year now, and we’re also hearing good things from users. The feature is relatively stable at this point, and people are often surprised how much time they can save just by getting up-to-date compile errors in milliseconds rather than seconds.

I haven’t really personally used incremental compilation with the LLVM backend, but all of the incremental test coverage in CI is now enabled for the LLVM backend, and I’ve had positive feedback from users, so it’s definitely worth giving a shot. As always, if you encounter bugs in incremental compilation, please report them if you can!

Thank you, and I hope you find this useful :)

Today, I merged a 30,000 line PR after two (arguably three) months of work. The goal of this branch was to rework the Zig compiler’s internal type resolution logic to a more logical and straightforward design. It’s a quite exciting change for me personally, because it allowed me to clean up a bunch of the compiler guts, but it also has some nice user-facing changes which you might be interested in!

For one thing, the Zig compiler is now lazier about analyzing the fields of types: if the type is never initialized, then there’s no need for Zig to care what that type “looks like”. This is important when you have a type which doubles as a namespace, a common pattern in modern Zig. For instance, when using std.Io.Writer, you don’t want the compiler to also pull in a bunch of code in std.Io! Here’s a straightforward example:

const Foo = struct { bad_field: @compileError("i am an evil field, muahaha"), const something = 123; }; comptime { _ = Foo.something; // Foo only used as a namespace }

Previously, this code emitted a compile error. Now, it compiles just fine, because Zig never actually looks at the @compileError call.

const Foo = struct {
    bad_field: @compileError("i am an evil field, muahaha"),
    const something = 123;
};
comptime {
    _ = Foo.something; // `Foo` only used as a namespace
}

Another improvement we’ve made is in the “dependency loop” experience. Anyone who has encountered a dependency loop compile error in Zig before knows that the error messages for them are entirely unhelpful—but that’s now changed! If you encounter one (which is also a bit less likely now than it used to be), you’ll get a detailed error message telling you exactly where the dependency loop comes from. Check it out:

const Foo = struct { inner: Bar }; const Bar = struct { x: u32 align(@alignOf(Foo)) }; comptime { _ = @as(Foo, undefined); }

$ zig build-obj repro.zig error: dependency loop with length 2 repro.zig:1:29: note: type 'repro.Foo' depends on type 'repro.Bar' for field declared here const Foo = struct { inner: Bar }; ^~~ repro.zig:2:44: note: type 'repro.Bar' depends on type 'repro.Foo' for alignment query here const Bar = struct { x: u32 align(@alignOf(Foo)) }; ^~~ note: eliminate any one of these dependencies to break the loop

Of course, dependency loops can get much more complicated than this, but in every case I’ve tested, the error message has had enough information to easily see what’s going on.

Additionally, this PR made big improvements to the Zig compiler’s “incremental compilation” feature. The short version is that it fixed a huge amount of known bugs, but in particular, “over-analysis” problems (where an incremental update did more work than should be necessary, sometimes by a big margin) should finally be all but eliminated—making incremental compilation significantly faster in many cases! If you’ve not already, consider trying out incremental compilation: it really is a lovely development experience. This is for sure the improvement which excites me the most, and a large part of what motivated this change to begin with.

There are a bunch more changes that come with this PR—dozens of bugfixes, some small language changes (mostly fairly niche), and compiler performance improvements. It’s far too much to list here, but if you’re interested in reading more about it, you can take a look at the PR on Codeberg—and of course, if you encounter any bugs, please do open an issue. Happy hacking!

const Foo = struct { inner: Bar };
const Bar = struct { x: u32 align(@alignOf(Foo)) };
comptime {
    _ = @as(Foo, undefined);
}

$ zig build-obj repro.zig
error: dependency loop with length 2
repro.zig:1:29: note: type 'repro.Foo' depends on type 'repro.Bar' for field declared here
const Foo = struct { inner: Bar };
^~~
repro.zig:2:44: note: type 'repro.Bar' depends on type 'repro.Foo' for alignment query here
const Bar = struct { x: u32 align(@alignOf(Foo)) };
^~~
note: eliminate any one of these dependencies to break the loop

As we approach the end of the 0.16.0 release cycle, Jacob has been hard at work, bringing std.Io.Evented up to speed with all the latest API changes:

io_uring implementation

Grand Central Dispatch implementation

Both of these are based on userspace stack switching, sometimes called “fibers”, “stackful coroutines”, or “green threads”.

They are now available to tinker with, by constructing one’s application using std.Io.Evented. They should be considered experimental because there is important followup work to be done before they can be used reliably and robustly:

better error handling

remove the logging

diagnose the unexpected performance degradation when using IoMode.evented for the compiler

a couple functions still unimplemented

more test coverage is needed

builtin function to tell you the maximum stack size of a given function to make these implementations practical to use when overcommit is off.

With those caveats in mind, it seems we are indeed reaching the Promised Land, where Zig code can have Io implementations effortlessly swapped out:

const std = @import("std");

pub fn main(init: std.process.Init.Minimal) !void { var debug_allocator: std.heap.DebugAllocator(.{}) = .init; const gpa = debug_allocator.allocator();

var threaded: std.Io.Threaded = .init(gpa, .{ .argv0 = .init(init.args), .environ = init.environ, }); defer threaded.deinit(); const io = threaded.io();

return app(io); }

fn app(io: std.Io) !void { try std.Io.File.stdout().writeStreamingAll(io, "Hello, World!\n"); }

$ strace ./hello_threaded execve("./hello_threaded", ["./hello_threaded"], 0x7ffc1da88b20 /* 98 vars */) = 0 ...

Swapping out only the I/O implementation:

const std = @import("std");

pub fn main(init: std.process.Init.Minimal) !void { var debug_allocator: std.heap.DebugAllocator(.{}) = .init; const gpa = debug_allocator.allocator();

var evented: std.Io.Evented = undefined; try evented.init(gpa, .{ .argv0 = .init(init.args), .environ = init.environ, .backing_allocator_needs_mutex = false, }); defer evented.deinit(); const io = evented.io();

return app(io); }

fn app(io: std.Io) !void { try std.Io.File.stdout().writeStreamingAll(io, "Hello, World!\n"); }

Key point here being that the app function is identical between those two snippets.

Moving beyond Hello World, the Zig compiler itself works fine using std.Io.Evented, both with io_uring and with GCD, but as mentioned above, there is a not-yet-diagnosed performance degradation when doing so.

Happy hacking,

Andrew

const std = @import("std");

pub fn main(init: std.process.Init.Minimal) !void {
    var debug_allocator: std.heap.DebugAllocator(.{}) = .init;
    const gpa = debug_allocator.allocator();

    var threaded: std.Io.Threaded = .init(gpa, .{
        .argv0 = .init(init.args),
        .environ = init.environ,
    });
    defer threaded.deinit();
    const io = threaded.io();

    return app(io);
}

fn app(io: std.Io) !void {
    try std.Io.File.stdout().writeStreamingAll(io, "Hello, World!\n");
}

const std = @import("std");

pub fn main(init: std.process.Init.Minimal) !void {
    var debug_allocator: std.heap.DebugAllocator(.{}) = .init;
    const gpa = debug_allocator.allocator();

    var evented: std.Io.Evented = undefined;
    try evented.init(gpa, .{
        .argv0 = .init(init.args),
        .environ = init.environ,
        .backing_allocator_needs_mutex = false,
    });
    defer evented.deinit();
    const io = evented.io();

    return app(io);
}

fn app(io: std.Io) !void {
    try std.Io.File.stdout().writeStreamingAll(io, "Hello, World!\n");
}

If you have a Zig project with dependencies, two big changes just landed which I think you will be interested to learn about.

Fetched packages are now stored locally in the zig-pkg directory of the project root (next to your build.zig file).

For example here are a few results from awebo after running zig build:

$ du -sh zig-pkg/* 13M freetype-2.14.1-alzUkTyBqgBwke4Jsot997WYSpl207Ij9oO-2QOvGrOi 20K opus-0.0.2-vuF-cMAkAADVsm707MYCtPmqmRs0gzg84Sz0qGbb5E3w 4.3M pulseaudio-16.1.1-9-mk_62MZkNwBaFwiZ7ZVrYRIf_3dTqqJR5PbMRCJzSuLw 5.2M uucode-0.1.0-ZZjBPvtWUACf5dqD_f9I37VGFsN24436CuceC5pTJ25n 728K vaxis-0.5.1-BWNV_AxECQCj3p4Hcv4U3Yo1WMUJ7Z2FUj0UkpuJGxQQ

It is highly recommended to add this directory to the project-local source control ignore file (e.g. .gitignore). However, by being outside of .zig-cache, it provides the possibility of distributing self-contained source tarballs, which contain all dependencies and therefore can be used to build offline, or for archival purposes.

Meanwhile, an additional copy of the dependency is cached globally. After filtering out all the unused files based on the paths filter, the contents are recompressed:

$ du -sh ~/.cache/zig/p/* 2.4M freetype-2.14.1-alzUkTyBqgBwke4Jsot997WYSpl207Ij9oO-2QOvGrOi.tar.gz 4.0K opus-0.0.2-vuF-cMAkAADVsm707MYCtPmqmRs0gzg84Sz0qGbb5E3w.tar.gz 636K pulseaudio-16.1.1-9-mk_62MZkNwBaFwiZ7ZVrYRIf_3dTqqJR5PbMRCJzSuLw.tar.gz 880K uucode-0.1.0-ZZjBPvtWUACf5dqD_f9I37VGFsN24436CuceC5pTJ25n.tar.gz 120K vaxis-0.5.1-BWNV_BFECQBbXeTeFd48uTJRjD5a-KD6kPuKanzzVB01.tar.gz

The motivation for this change is to make it easier to tinker. Go ahead and edit those files, see what happens. Swap out your package directory with a git clone. Grep your dependencies all together. Configure your IDE to auto-complete based on the zig-pkg directory. Run baobab on your dependency tree. Furthermore, by having the global cache have compressed files instead makes it easier to share that cached data between computers. In the future, it is planned to support peer-to-peer torrenting of dependency trees. By recompressing packages into a canonical form, this will allow peers to share Zig packages with minimal bandwidth. I love this idea because it simultaneously provides resilience to network outages, as well as a popularity contest. Find out which open source packages are popular based on number of seeders!

$ du -sh zig-pkg/*
13M freetype-2.14.1-alzUkTyBqgBwke4Jsot997WYSpl207Ij9oO-2QOvGrOi
20K opus-0.0.2-vuF-cMAkAADVsm707MYCtPmqmRs0gzg84Sz0qGbb5E3w
4.3M pulseaudio-16.1.1-9-mk_62MZkNwBaFwiZ7ZVrYRIf_3dTqqJR5PbMRCJzSuLw
5.2M uucode-0.1.0-ZZjBPvtWUACf5dqD_f9I37VGFsN24436CuceC5pTJ25n
728K vaxis-0.5.1-BWNV_AxECQCj3p4Hcv4U3Yo1WMUJ7Z2FUj0UkpuJGxQQ

$ du -sh ~/.cache/zig/p/*
2.4M freetype-2.14.1-alzUkTyBqgBwke4Jsot997WYSpl207Ij9oO-2QOvGrOi.tar.gz
4.0K opus-0.0.2-vuF-cMAkAADVsm707MYCtPmqmRs0gzg84Sz0qGbb5E3w.tar.gz
636K pulseaudio-16.1.1-9-mk_62MZkNwBaFwiZ7ZVrYRIf_3dTqqJR5PbMRCJzSuLw.tar.gz
880K uucode-0.1.0-ZZjBPvtWUACf5dqD_f9I37VGFsN24436CuceC5pTJ25n.tar.gz
120K vaxis-0.5.1-BWNV_BFECQBbXeTeFd48uTJRjD5a-KD6kPuKanzzVB01.tar.gz

The second change here is the addition of the --fork flag to zig build.

In retrospect, it seems so obvious, I don’t know why I didn’t think of it since the beginning. It looks like this:

zig build --fork=[path]

This is a project override option. Given a path to a source checkout of a project, all packages matching that project across the entire dependency tree will be overridden.

Thanks to the fact that package content hashes include name and fingerprint, this resolves before the package is potentially fetched.

This is an easy way to temporarily use one or more forks which are in entirely separate directories. You can iterate on your entire dependency tree until everything is working, while using comfortably the development environment and source control of the dependency projects.

The fact that it is a CLI flag makes it appropriately ephemeral. The moment you drop the flags, you’re back to using your pristine, fetched dependency tree.

If the project does not match, an error occurs, preventing confusion:

$ zig build --fork=/home/andy/dev/mime error: fork /home/andy/dev/mime matched no mime packages $

If the project does match, you get a reminder that you are using a fork, preventing confusion:

$ zig build --fork=/home/andy/dev/dvui info: fork /home/andy/dev/dvui matched 1 (dvui) packages ...

This functionality is intended to enhance the workflow of dealing with ecosystem breakage. I already tried it a bit and found it to be quite pleasant to work with. The new workflow goes like this:

Fail to build from source due to ecosystem breakage.

Tinker with --fork until your project works again. During this time you can use the actual upstream source control, test suite, zig build test --watch -fincremental, etc.

Now you have a new option: be selfish and just keep working on your own stuff, or you can proceed to submit your patches upstream.

…and you can probably skip the step where you switch your build.zig.zon to your fork unless you expect upstream to take a long time to merge your fixes.

zig build --fork=[路径]

$ zig build --fork=/home/andy/dev/mime
error: fork /home/andy/dev/mime matched no mime packages
$

$ zig build --fork=/home/andy/dev/dvui
info: fork /home/andy/dev/dvui matched 1 (dvui) packages
...

The Windows operating system provides a large ABI surface area for doing things in the kernel. However, not all ABIs are created equally. As Casey Muratori points out in his lecture, The Only Unbreakable Law, the organizational structure of software development teams has a direct impact on the structure of the software they produce.

The DLLs on Windows are organized into a heirarchy, with some of the APIs being high-level wrappers around lower-level ones. For example, whenever you call functions of kernel32.dll, ultimately, the actual work is done by ntdll.dll. You can observe this directly by using ProcMon.exe and examining stack traces.

What we’ve learned empirically is that the ntdll APIs are generally well-engineered, reasonable, and powerful, but the kernel32 wrappers introduce unnecessary heap allocations, additional failure modes, unintentional CPU usage, and bloat.

This is why the Zig standard library policy is to Prefer the Native API over Win32. We’re not quite there yet - we have plenty of calls into kernel32 remaining - but we’ve taken great strides recently. I’ll give you two examples.

According to the official documentation, Windows does not have a straightforward way to get random bytes.

Many projects including Chromium, boringssl, Firefox, and Rust call SystemFunction036 from advapi32.dll because it worked on versions older than Windows 8.

Unfortunately, starting with Windows 8, the first time you call this function, it dynamically loads bcryptprimitives.dll and calls ProcessPrng. If loading the DLL fails (for example due to an overloaded system, which we have observed on Zig CI several times), it returns error 38 (from a function that has void return type and is documented to never fail).

The first thing ProcessPrng does is heap allocate a small, constant number of bytes. If this fails it returns NO_MEMORY in a BOOL (documented behavior is to never fail, and always return TRUE).

bcryptprimitives.dll apparently also runs a test suite every time you load it.

All that ProcessPrng is really doing is NtOpenFile on "\Device\CNG" and reading 48 bytes with NtDeviceIoControlFile to get a seed, and then initializing a per-CPU AES-based CSPRNG.

So the dependency on bcryptprimitives.dll and advapi32.dll can both be avoided, and the nondeterministic failure and latencies on first RNG read can also be avoided.

ReadFile looks like this:

pub extern "kernel32" fn ReadFile( hFile: HANDLE, lpBuffer: LPVOID, nNumberOfBytesToRead: DWORD, lpNumberOfBytesRead: ?*DWORD, lpOverlapped: ?*OVERLAPPED, ) callconv(.winapi) BOOL;

NtReadFile looks like this:

pub extern "ntdll" fn NtReadFile( FileHandle: HANDLE, Event: ?HANDLE, ApcRoutine: ?*const IO_APC_ROUTINE, ApcContext: ?*anyopaque, IoStatusBlock: *IO_STATUS_BLOCK, Buffer: *anyopaque, Length: ULONG, ByteOffset: ?*const LARGE_INTEGER, Key: ?*const ULONG, ) callconv(.winapi) NTSTATUS;

As a reminder, the above function is implemented by calling the below function.

Already we can see some nice things about using the lower level API. For instance, the real API simply gives us the error code as the return value, while the kernel32 wrapper hides the status code somewhere, returns a BOOL and then requires you to call GetLastError to find out what went wrong. Imagine! Returning a value from a function 🌈

Furthermore, OVERLAPPED is a fake type. The Windows kernel doesn’t actually know or care about it at all! The actual primitives here are events, APCs, and IO_STATUS_BLOCK.

If you have a synchronous file handle, then Event and ApcRoutine must be null. You get the answer in the IO_STATUS_BLOCK immediately. If you pass an APC routine here then some old bitrotted 32-bit code runs and you get garbage results.

On the other hand if you have an asynchronous file handle, then you need to either use an Event or an ApcRoutine. kernel32.dll uses events, which means that it’s doing extra, unnecessary resource allocation and management just to read from a file. Instead, Zig now passes an APC routine and then calls NtDelayExecution. This integrates seamlessly with cancelation, making it possible to cancel tasks while they perform file I/O, regardless of whether the file was opened in synchronous mode or asynchronous mode.

For a deeper dive into this topic, please refer to this issue:

Windows: Prefer the Native API over Win32

pub extern "kernel32" fn ReadFile(
    hFile: HANDLE,
    lpBuffer: LPVOID,
    nNumberOfBytesToRead: DWORD,
    lpNumberOfBytesRead: ?*DWORD,
    lpOverlapped: ?*OVERLAPPED,
) callconv(.winapi) BOOL;

pub extern "ntdll" fn NtReadFile(
    FileHandle: HANDLE,
    Event: ?HANDLE,
    ApcRoutine: ?*const IO_APC_ROUTINE,
    ApcContext: ?*anyopaque,
    IoStatusBlock: *IO_STATUS_BLOCK,
    Buffer: *anyopaque,
    Length: ULONG,
    ByteOffset: ?*const LARGE_INTEGER,
    Key: ?*const ULONG,
) callconv(.winapi) NTSTATUS;

Over the past month or so, several enterprising contributors have taken an interest in the zig libc subproject. The idea here is to incrementally delete redundant code, by providing libc functions as Zig standard library wrappers rather than as vendored C source files. In many cases, these functions are one-to-one mappings, such as memcpy or atan2, or trivially wrap a generic function, like strnlen:

fn strnlen(str: [*:0]const c_char, max: usize) callconv(.c) usize { return std.mem.findScalar(u8, @ptrCast(str[0..max]), 0) orelse max; }

So far, roughly 250 C source files have been deleted from the Zig repository, with 2032 remaining.

With each function that makes the transition, Zig gains independence from third party projects and from the C programming language, compilation speed improves, Zig’s installation size is simplified and reduced, and user applications which statically link libc enjoy reduced binary size.

Additionally, a recent enhancement now makes zig libc share the Zig Compilation Unit with other Zig code rather than being a separate static archive, linked together later. This is one of the advantages of Zig having an integrated compiler and linker. When the exported libc functions share the ZCU, redundant code is eliminated because functions can be optimized together. It’s kind of like enabling LTO (Link-Time Optimization) across the libc boundary, except it’s done properly in the frontend instead of too late, in the linker.

Furthermore, when this work is combined with the recent std.Io changes, there is potential for users to seamlessly control how libc performs I/O - for example forcing all calls to read and write to participate in an io_uring event loop, even though that code was not written with such use case in mind. Or, resource leak detection could be enabled for third-party C code. For now this is only a vaporware idea which has not been experimented with, but the idea intrigues me.

Big thanks to Szabolcs Nagy for libc-test. This project has been a huge help in making sure that we don’t regress any math functions.

As a reminder to our users, now that Zig is transitioning to being the static libc provider, if you encounter issues with the musl, mingw-w64, or wasi-libc libc functionality provided by Zig, please file bug reports in Zig first so we don’t annoy maintainers for bugs that are in Zig, and no longer vendored by independent libc implementation projects.

fn strnlen(str: [*:0]const c_char, max: usize) callconv(.c) usize {
    return std.mem.findScalar(u8, @ptrCast(str[0..max]), 0) orelse max;
}